Shopify
Stripe
Slack
Notion
Google Workspace
Confluence
How Do You Connect a Knowledge Base to an AI Agent? (2026)
An AI agent answers accurately only when it can read your own help center, policies and docs. Here is which sources you can connect, how the agent reads each one, and the order to connect them in.
Key takeaways
- Connecting a knowledge base to an AI agent means adding each source, letting the platform index it, then linking it to the agent with an all-documents or selected scope.
- Macha accepts uploaded files up to 20 MB each, 10 at a time, and crawls a website up to 200 pages, while Google Docs, Notion and Confluence are read live.
- Macha loads spreadsheets of 2,000 rows or fewer straight into the agent's context, while larger files and websites are chunked into embeddings and searched.
- Uploaded files and crawled websites in Macha do not refresh on their own, so an edited document needs a re-upload and a changed site needs a Resync.
- Macha bills per ticket from $299 a month for 750 tickets, so the size of a connected knowledge base never changes what a conversation costs.
To connect a knowledge base to an AI agent, you add each source (help center, uploaded files, website, Notion, Confluence or Google Docs), let the platform index it, then link it to the agent with a scope of all documents or selected ones. In Macha that takes minutes: files upload 10 at a time at up to 20 MB each, a website crawl indexes up to 200 pages, and live sources like Google Docs are read fresh on every question.
| Source | How the agent reads it | How it stays current |
|---|---|---|
| Zendesk Help Center | Indexed and searched | Live webhook sync |
| Uploaded files (PDF, DOCX, TXT, CSV, XLSX) | Indexed; spreadsheets of 2,000 rows or fewer go straight into the agent's context | Delete and re-upload |
| Website (crawl up to 200 pages) or single page | Indexed and searched | Resync button |
| Google Docs, Notion, Confluence | Read live through the connector | Always the current version |
| Studies results (past tickets) | Exported as a searchable source | Export again after a new run |
Why does grounding an agent in your knowledge matter?
A raw language model knows a lot about the world and nothing about your product, policies, or edge cases. Point an agent at your knowledge base and three things change:
- Accuracy: it answers from your documented truth, not a plausible guess.
- Fewer hallucinations: when the answer is in your content, it uses it; when it isn't, a well-built agent escalates instead of inventing.
- Current answers: update an article and the agent's answers update with it.
How does an AI agent search a knowledge base?
There are three modes, and which one a source gets depends on its size and type. Large files, PDFs, DOCX files and website pages are indexed: each document is split into chunks, turned into embeddings (vector representations for semantic search), and stored. When a customer asks something, the agent calls a search tool to find the most relevant chunks, then pulls the full section it needs. Small spreadsheets (2,000 rows or fewer) skip search entirely and are injected into the agent's context, so every row is always available. Connector sources such as Google Docs and Notion are read live: the agent fetches the current document each time.
One agent can mix all three modes.
What knowledge sources can you connect?
With a platform like Macha, an agent can draw on a mix of sources:
- Zendesk Help Center: articles sync over a live webhook, so a newly published article reaches the agent without a manual refresh.
- Notion: pages are read live.
- Confluence: pages are read live, and there's a ready-made template that searches Confluence for each new ticket and posts the relevant links as an internal note.
- Google Docs: documents are read live; linking one auto-adds the Google read tools to the agent.
- Uploaded files: PDF, DOCX, TXT, CSV, XLSX, up to 10 files at a time and 20 MB each.
- Your website: crawl a site (up to 200 pages) or index a single page.
- Past tickets: results from a Studies run can be exported as a searchable source, one document per row.
How do you connect your knowledge base, step by step?
- Connect the source. For Zendesk, connecting the Help Center sets up the sync. For Notion, Confluence, or Google, authorize the connection and pick what to include.
- Let it index. Uploaded files usually index in a few seconds to a couple of minutes; a site crawl usually finishes in a few minutes and shows a Ready status.
- Link the source to your agent. Choose All documents (new documents in the source are picked up automatically) or Selected (only the documents you tick), so each agent sees exactly the knowledge it should.
- Test it. Ask the agent questions you know the answers to, and check it pulls the right content.
- Keep it fresh. The Help Center and live connectors stay current on their own; hit Resync on a website source after the site changes, and re-upload a file after you edit the original.
On cost, Macha bills per ticket: one conversation, charged once, however many articles the agent reads or steps it takes to answer. Indexing and storing your knowledge isn't billed separately, so a big knowledge base costs nothing extra, and a well-grounded answer costs the same as a vague one. The one plan starts at $299/mo for 750 tickets, about $0.40 each, with setup and monitoring by the Macha team included. $50 of free usage, no credit card required.
What should you connect first?
A pragmatic order:
- Your Help Center: it syncs itself and already covers your most common questions.
- Your top 20 articles by ticket volume: if they aren't in the Help Center yet, add them; this is where most resolutions come from.
- Policies (refunds, shipping, billing): the rules the agent needs to answer correctly and consistently.
- Internal docs (Notion/Confluence) for the questions your public content doesn't cover.
Then watch what the agent gets asked that it can't answer, and fill those gaps. The questions customers ask tell you exactly what to write next. If your help center is thin, Macha's AI Knowledge Builder drafts articles from your historical tickets.
What are the best practices for a knowledge base an agent reads?
- Curate, don't dump. Connect your good content. Garbage in, confident-wrong out.
- Keep it current. Retire stale articles; the agent will faithfully repeat whatever you leave in.
- Have a single source of truth. Avoid three conflicting versions of the same policy.
- Scope per agent. A refund agent doesn't need your engineering wiki.
- Write for retrieval. Clear titles and focused articles get found more reliably.
What goes wrong most often?
- Thin or outdated knowledge leads to vague or wrong answers. Fix the content first.
- No escalation path means the agent guesses when it shouldn't. Make sure it hands off when unsure.
- Over-broad sourcing surfaces irrelevant docs. Scope tightly.
- Stale snapshots: an uploaded file or crawled site doesn't refresh itself, so an edited original keeps giving the old answer until you re-upload or resync.
Frequently asked questions
How does an AI agent use my knowledge base? It indexes your content, then searches it to answer each question from the most relevant passages, and can pull a full document when it needs detail.
What knowledge sources can I connect? Your Zendesk Help Center, Notion, Confluence, Google Docs, uploaded files (PDF, DOCX, TXT, CSV, XLSX), a crawled website of up to 200 pages, and exported past-ticket analysis from Studies.
Does the agent stay up to date when I change an article? For the Zendesk Help Center and live connectors, yes. Websites need a Resync, and uploaded files need to be re-uploaded after you edit them.
Will the agent make things up? A well-built one answers from your knowledge and escalates when the answer isn't there. Grounding plus a clear escalation path is what keeps it honest.
Does connecting a big knowledge base cost a lot? No. Indexing isn't billed separately, and billing is per ticket, so the size of your knowledge base never changes what a conversation costs.
What should you do next?
Put anything that changes weekly behind a live connector, then test with questions you already know the answers to. Macha runs on Zendesk, Freshdesk, Gorgias, Front, HubSpot or Intercom, so the same knowledge setup works whichever desk your team uses.
Ground an agent in your own knowledge: connect your help center and docs, and build an agent that answers from them. $50 of free usage, no credit card required. Start a free trial.
